International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
Face Detection Identification and Tracking by
PRDIT Algorithm using Image Database for
Crime Investigation
V. S. Manjula
Lt. Dr. S. Santhosh Baboo
Asst. Professor,
Dept. of Computer Application,
St. Joseph’ College,
Chennai- 600 122, INDIA
Reader, P.G & Research,
Dept. of Computer Application,
D.G. Vaishnav College,
Chennai-106. INDIA
ABSTRACT
In general, the field of face recognition has lots of research
that have put interest in order to detect the face and to identify
it and also to track it. Many researchers have concentrated on
the face identification and detection problem by various
approaches. The proposed approach is further very useful and
helpful in very huge real time application. Thus the Face
Detection, Identification and Tracking mechanism which is
proposed here is used to detect the faces in videos in the real
time application by using the PRDIT (Proposed Rectangular
Detection Identification and tracking) algorithm. Thus the
proposed mechanism is very help full in identifying individual
persons who are been involved in the action of robbery,
murder cases and terror activities. Although in face
recognition the algorithm used is of histogram equalization
combined with Back propagation neural network in which we
recognize an unknown test image by comparing it with the
known training set images that are been stored in the database.
Also the proposed approach uses skin color extraction as a
parameter for face detection. A multi linear training and
rectangular face feature extraction are done for training,
identifying, detecting and tracking. Thus the proposed
technique is very useful in identify a single person from a
group of faces. Thus the proposed technique is well suited for
all kinds of faces which have a specific complexion varying
under certain range. Also we have taken a real life example
and simulated the algorithms in IDL Tool successfully.
Keywords:
PRDIT, histogram equalization, rectangular
features.
1. INTRODUCTION
The face is our primary focus of attention in social life
playing an important role in conveying identity and emotions.
Detection and recognisation of face is an important research
in the area of computer vision [1]. The process of face
detecting is a challenge and toughest process because of
different facial expression, races, backgrounds, illumination,
overlapping, low brightness make the face detection process
as more complicated one.
Computational models of face recognition are interesting
because they can contribute not only to theoretical knowledge
but also to practical applications. Computers that detect and
recognize faces could be applied to a wide variety of tasks
including criminal identification, security system, image and
film processing, identity verification, tagging purposes and
human-computer interaction. Unfortunately, developing a
computational model of face detection and recognition is quite
difficult because faces are complex, multidimensional and
meaningful visual stimuli [2].
Thus the Face detections are used in many places now a days
especially on websites hosting like picassa and face book. The
automatically tagging feature adds a new dimension in order
to share pictures among the people who are in the picture and
also gives idea to other people about who the person in the
image. Here we have studied and implemented a pretty simple
but very effective face detection algorithm which takes human
skin colour into account for detecting and tracking the
face[3][4].
Also here we propose a new approach which is based on the
multi linear training and rectangular face feature extraction.
Detecting, training, tracking and identification are the major
steps of the proposed technique. The main aim, which we
believe we have reached, was to develop a method of face
recognition that is of fast, robust, reasonably simple and
accurate with a relatively simple and easy to understand
algorithms and techniques [4].
2. PERSPECTIVE OF OUR WORK
The proposed approach is simple, fast and accurate which is
been applied together as a single algorithm to provide better
results under complex circumstances like face position,
luminance variation etc. Each of these algorithms are been
discussed one by one below. Thus the proposed approach
handle changes on the face image like lighting, complexity in
the background, multiple faces in the image. Thus the
proposed approach makes an improvement in the detection
results rather than the other detectors. The more challenging
function in the detectors is to handle the poses. Different type
of pose makes conflict in detecting a particular personality
and our proposed approaches overcome these drawbacks [1].
Next the various method of tracking approach gets confused
in the beginning stage and we use a tracker algorithm for the
detection and it avoids the above mentioned problem. The
information for the detecting process incorporates with the
parameter and reproduces the information itself. Hence these
algorithms always find faces in the frames even though the
frame based detectors gets fails. Thus the knowledge of
training can understand new faces that are entered in the
training and it is always ready to integrate in the updating
process [7].
41
International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
2.1 Training Phase:
The training process is done by applying a back-propagation
neural network. The input will be of face features along with
that the individual users’ names are fed into the system for
processing. In order to extract the skin color from the face
images we apply a histogram equalization process. Thus the
trained data hold individual users skin color and some extra
facial features. Also here we introduce a new proposed
rectangle based approach for detecting the face [8]. In this
approach the human face images are been splitted into various
rectangle portions i.e. of with different shapes and size also
with their different features are also been saved along with the
training samples in databases for the later verification. Figure
1.explains about the training procedure of our proposed
approach.
Figure 1.Trainning Module
For training the database 100 input face images with
different features are given as input to the training system.
The input images are taken from GTAV Face Database for
training the input datasets [10]. The face and non-face block
weights are calculated as negative and positive values for the
neural network. Also our proposed neural network system is
made up with 9 hidden layers.
Pseudo code of the Network:
o
o
o
-
-
-
Weight as the random values with the range -1.0 to 1.0
Input pattern as binary value for input layer of the neural
network.
Neuron Process:
Multiply neuron leading connection weight values with
previous neuron output values.
Add all values.
Apply activation function to generate output.
Output layer is reached to stop this process.
- Output pattern is checked with the target pattern for
calculate
the value of mean square error.
-Weight (old) + Learning Rate * Output Error * Output
(Neuron i) * Output (Neuron i + 1) * (1 – Output
(Neuron i + 1))
Go to the Initial step.
The algorithm should be end if the output of the all
pattern should be matched with the pattern of the target
[9].
Train1 Ii (Ni)
Where, Ii  Intensity,
Ni Different angle image for single person ,
Train1 Multi linear training
2.2 Preprocessing of the Image
Face images are given as input to the proposed system which
extracts multi features of the face. Also various face regions
are taken under some conditions e.g. background, light and
distance between the camera and face makes the different
appearance of the face in the image. For getting the scale
variance property, the input images where resized as pyramid
image. Also we partition the image as 21 x 21 pixel blocks
which is present in the sub-sampled image and then normalize
the blocks of the image to unit variance and zero means.
We take an RGB color image as input
(Ri, Gi, Bi) Є [0.1]3
And produce a grayscale image as output
T Є [0.1]
To avoid gamma correction issues.
2.2.1. Mathematical approach
Apply the same method individually to the RGB image i.e.,
red, green and blue components in the color model. Applying
the similar color method to the RGB image may produce the
different color balance in the images. Since the color channels
changes their algorithm for the getting the relative
distribution. First of all the image is changed to another space
color like HSV/HSL or Lab color space then we apply the
proposed algorithm to value channels or luminance which
does not affect the image hue and saturation.
Let us take the image of discrete grayscale and consider ni be
the gray level occurrence number i.
42
International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
The probability of an occurrence of a pixel of level i in the
image is
P (xi) = ni / n
Where, i Є 0… L - 1
L  gray level total number in the image.
n pixels total number of the image
P in fact the image's histogram and Normalized to [0, 1]
x an occurrence of a pixel in the image
C (i) =Σi j=0 P (xj)
Where C the cumulative distribution function
corresponding to p,
To produce the original image of level y on the original image
level on x we have to transform the form .i.e Y’s cumulative
probability function be the linearized in the range of value.
The transformation is defined by:
Yi= T (xi) = c (i)
Yi  linearized transformation of pixels
Note that the T plots the levels between the domain
of 0…1. to plot the values in the original domain The
following transformation should be applied to the results.
y’i=yi . (Max – min) + Min
’
y i equalized image
MaxMaximam value pixel in the image
Min  Minmam value pixel in the image
3. PRDIT APPROACH
Proposed Rectangular Detection Identification and tracking
(PRDIT) technique is an extension of RDIT that uses
histogram equalization process and skin color extraction and
neural network with the proficient of learning the interactions
of the multiple factors like different viewpoints, different
lighting conditions, different expressions etc. In the face
detection process skin color plays an important role when
extracting the information in the area of face [8]. That
obtained color information is used to make the task easier.
The process of identifying the face localization is very hard to
achieve but the color information makes it easier in the
complex environment. Based on the various model of color
space like YCbCr, HSV and YIQ the researchers offer the
many techniques for detecting the skin information.
Generally no one have the same skin information. The color
of the skin is differed in each person. It is arranged in the
small portion of area in CbCr plane [9]. This model is the
strong as compared with the other model having the skin
information. The people in the different areas like Europe,
Africa and Asia has the different type of skin in formation. So
we implement the YCbCr color model to detect the color of
the skin [10].
3.1 Detection of the Face
We randomly provide the image input to find whether the
human being face is present or not in the input photos or
videos. If the face is present, it return the location and the
degree of the space between each face.
GRectsi=GRectf (Ii>equalized value)
SRectsi =GRectsi (w*h) / ni
If Ii (SRectsi) = Train2
This rectangle is in face
Else
This rectangle not in face
Where
GRectsiFace rectangle Group image
GRectfFull input Group mage Rectangle
SRectsi Small rectangle Group image
3.2 Identification of the Face:
All face images are divided into blocks of triangles which are
providing to the neural network. Some of the face candidate
blocks will be in the output of the networks. To eliminate the
false detection blocks we present the face verification
methods. The basic idea behind in this concept is considering
the edge point distribution in facial features. By using the
method of the blocks of the each candidates the edge map
detecting is done. Then the proposed algorithm stores the
features of the face rectangle for the various images. The
images which are stored in the databases will be matched with
the any of the features of the input images.
If Ii (GRectsi) = Train1
Face Identified who is that person
STRect= GRectsi
Else
No one Identified
Where
STRectStore the Identified face rectangle
3.3 Face Tracking & Recognisation:
In the active scenario, the tracking is the more important
process and it follows the sequence of process. Also the
process of tracking is divided into two categories.
1) Head tracking
2) Facial feature tracking
Head tracking methods is color, region and shape
based one and the feature tracking method has trackers for
each and every feature and outlines the points like mouth,
eyes for the tracking process. For this process of predicting
and updating Condensation filter and Kalman filter are used.
The grouping of the head tracking and facial feature tracking
along with the concepts of filter reduce the overall problem.
The framework mainly focused to watch out the multiple
faces at same time. By using the Kalman filters the process of
tracking of multiple people were performed.
If Ii (NFRectsi) = STRect
Identified person was present in the frame
Else
Identified person was not present in the frame
Where
NF Next Frame of Group image (applied
preprocessing and Face Detection)
NFRectsi Next Frame Face rectangles
4. RESULTS AND DISCUSSION:
To deliver the effective system we take the various color
images which includes many faces with the numerous size
and have the dissimilar lighting circumstance and different
expressions. The color of the skin may be varied because of
the different lighting effect. To minimize the lighting effects
we include the gray methods which carry out the damages in
the light effects. Then the color components of RGB color i.e.,
Red, Green and Blue nonlinearly changed to the YCbCr color
space.
43
International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
Fig.2.1 a
Fig.2.2 a
Fig.2.3 a
Fig.2.1
Fig.2.1
Fig.2.1
Fig.2.1
b
c
d
e
Fig.2.2
Fig.2.2
Fig.2.2
Fig.
b
c
d
2.2 e
Fig.2.3
Fig.2.3
Fig.2.3
Fig.
b
c
d
2.3 e
Figure 4.1, 4.2, 4.3 Comparison of (a)original (b) Grey Scale (c) Histogram equalization Image (d) Skin Color
Extraction (e)
Thus the preferred threshold is [Cb1, Cb2] & [Cr1, Cr2]. The
value of the [Cb, Cr] presents with in the threshold limit is
considered to have a good skin tone of pixels. If the threshold
limit is higher then it is considered to be a dull skin with less
pixel value .Thus the proposed algorithm presents is an
outstanding performance by showing the evidence above.
Hence we perform our experiments with different par
parameters and features. Figure 2.4 a explains about the
comparison between the number of trained input vs. accuracy
of result. Figure 2.4 b explains about the comparison between
number of trained input vs. speed of detection which is been
compared with the existing PCA, RDIT and PRDIT
approaches.
Figure 2.4 a Trained input vs. accuracy
44
International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
Figure 2.4 b Trained input vs. speed of detection
The features of test image were also extracted in the same
fashion and were compared with the trained database. The
recognition was successful on 95% of the occasion. The
recognition algorithm was same as PRDIT and it involved the
training approach.
Hence either way it detected a face either true or false. The
quantitative success rates are provided below. Also a
comparison between PCA, RDIT and PRDIT algorithm is
done to highlight the advantages of this algorithm over PCA
and RDIT.
PCA
RDIT
PRDIT
Requires full frontal display
Requires full frontal display with no
lightning conditions
Works well with different viewpoints,
expressions and lighting conditions
Each face is a single entity in the database
Training is done on faces of same person
with different features
Training is done on faces of same
person with different features
Recognisation is done by one to one
matching
Recognisation is done by different pose of
same person
Recognisation is done from the group of
different users.
Table 1 comparison between PCA, RDIT and PRDIT
45
International Journal of Computer Applications (0975 – 8887)
Volume 38– No.10, January 2012
Thus these proposed techniques works well under any robust
conditions like complex background and also with different
face positions. These algorithms give different rates of
accuracy under different conditions as experimentally
observed.
A test e shown in figure 2.5a was taken on the training of
different of different types of technique versus accuracy in
percentage were highlighted.
Figure 2.5 a Different algorithm vs. accuracy
Figure 2.5 b Different algorithm vs. Error Rates
The experiment was conducted again and again which
provided a 100% result. Also the experiment was proved in
Figure 2.5 b that the proposed algorithm has a tolerable Error
rates.
5. CONCLUSION
PRDIT is useful for face Detection, Identification and
tracking mechanism and also used to detect the faces in videos
in the real time. Thus the proposed technique is very useful in
identify a single person from a group of faces in a set of video
frames. Thus the proposed technique is well suited for all
kinds of faces. Thus the face recognition, detection and
tracking algorithms were implemented and tested with
different types of images under varying conditions. The
RDIT, PCA and PRDIT success rates were given for face
detection and the success rate was different for different
images depending on the external factors. The overall success
rate was 95%.
6. REFERENCES
[1] Chiang.C.C et al., “A novel method for detecting lips
eyes and faces in real time”, Real Time Imaging-2003
[2] Garcia.C & Tziritas.G “Face detection using quantized
skin color regions merging and wavelet packet analysis”,
IEEE Trans. Multimedia-1999
[4] HsuR.L. & Jain.A.K and Abdel Mottaleb.M “Face
detection in color images”, IEEE Trans-2002
[5] Kruppa.H & Schiele.B “Using Local Context to Improve
Face Detection”, In Proc. of the British Machine Vision
Conference (BMVC), Norwich-2003
[6] Mostafa.L & Abdelazeem.S ”Face Detection based on
Skin Color using Neural Networks”, in Proc. of 1 st
ICGST Intl Conference on Graphics, Vision and Image
Processing GVIP, Egypt-2005
[7] Pitas.I & Karasaridis.A “Multichannel transforms for
signal image processing” IEEE Trans Image Processing1996
[8] Rizzi.A & Marini.D and Gatta.C “Color correction
between gray world and white patch”, The human Vision
and Electronic Imaging Conf-2002
[9] Viola.P & Jones.M.J “Robust real time face detection”,
Intl Journal of Computer Vision-2004
[10] Wang.H & Chang.S.F “A highly efficient system for
automatic face region detection in MPEG videos”, IEEE
Trans-Circuit Systems for Video Technology in 1997
[11] Yepeng Guan & Lin Yang “unsupervised face detection
based on skin color and geometric information”
[3] 3. Guetta.A & Rajagopal.s and Pare.M “Face Detection
EE368 Final Project”- 2003
46